2026-11-09 · America/Los_Angeles · 论文 · #18

CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse

Large language models reinvent the wheel all the time. They conflate task-solving with the execution of the solutions, which forces them to solve every problem from scratch. We show that once a task has been reasoned through, the application of its solution to a new problem instance can be done very efficiently and with little involvement from language models. We introduce agentic reasoning caching, a method for constructing robust CPU-executable agentic reasoning caches (ARCs) from LLM reasoning traces. ARCs show on-par performance with frontier LLMs across a range of difficult mathematical…

热度 32.4 / 100；排名与评分保留该期记录。